Local agent workflow

ChatGPT Delegate for Codex

Ask Codex to hand off deep research, PRD writing, comparison, and quick web lookup tasks to ChatGPT Web. Codex handles orchestration; ChatGPT does the long-form work; a local MCP connector saves the result as Markdown.

Use from Codex View on GitHub
ChatGPT Delegate data flow from Codex to Chrome, ChatGPT, MCP, and local Markdown

Save Codex tokens

Codex prepares the task and waits for status. It does not generate or read the full body unless you explicitly ask.

Use ChatGPT Web

The actual research, thinking, writing, and browsing happen in the signed-in ChatGPT Web/App surface.

Keep local files

Results land under artifacts/chatgpt-delegation/<task_id>/result.md with metadata.

Task Profiles

research_pro

For deep research, community methods, best practices, and long reports with sources.

thinking_high

For design work, PRDs, planning, comparisons, structured writing, and strategy drafts.

instant_web

For quick lookup, latest facts, confirmation checks, and short source-backed answers.

Data Flow

  1. In Codex, say something like 让 ChatGPT 调研一下 xxx, 让 ChatGPT 写一个 PRD, or 让 ChatGPT 查一下最新情况.
  2. Codex runs chatgpt-delegate prepare, which creates task.json, prompt.md, and a target result path.
  3. Codex starts the local MCP server and a temporary HTTPS tunnel if ChatGPT needs to reach the connector.
  4. Codex opens ChatGPT Web through browser automation and sends the generated prompt to the ChatGPT Delegate connector app.
  5. ChatGPT performs the research, thinking, writing, or web lookup in ChatGPT Web.
  6. ChatGPT calls save_task_result on the MCP connector, writing result.md and result.json.
  7. Codex polls task status and returns only the profile, status, byte count, and local file path by default.

Use From Codex

Install a Codex skill named chatgpt-delegate and expose a local MCP server with the tool contract below. Then use natural language in Codex:

让 ChatGPT 调研一下 100M 上下文大模型的社区方法
让 ChatGPT 设计一个 ChatGPT x Codex 的 PRD
让 ChatGPT 查一下 xurl OAuth2 当前最佳实践

The important rule is the execution split: Codex orchestrates the browser and file checks; ChatGPT produces the substantive Markdown body.

MCP Tool Contract

save_task_result(task_id, title, markdown, summary=None, overwrite=True)
list_results(limit=20)
read_result(task_id)
connector_status()

The status tool should make the boundary explicit: codex_execution=false, openai_api=false, and private_chatgpt_api=false.

This workflow does not read ChatGPT session tokens, does not call private ChatGPT APIs, and does not require an OpenAI API key for the delegated body generation.